[TimelordUK]

R13 slice 4, third part. WHERE's IS NULL, IS NOT NULL and LIKE now go through the value evaluator, so every WHERE operator now delegates. What's left of slice 4 (AND / OR / NOT / CASE) waits on slice 5, method calls.

Closes P55: a LIKE pattern in WHERE was read as a regex

WHERE built its matcher by replacing % and _ and escaping nothing else, so any regex syntax in a pattern took effect:

• LIKE 'a.c' matched abc too (so LIKE '%.csv' matches xcsv)
• LIKE '[a]bc' matched abc but not the literal [a]bc
• LIKE 'a(b%' failed the whole query: regex parse error … unclosed group
• % stopped at a newline

The bug was in both the cached path (EvaluationContext) and the uncached fallback.

Commits

1. IS NULL / IS NOT NULL delegate. A true no-op: the parser always puts a Null literal on the right. Also deletes the IS, IS NOT and == operator arms, which the parser never produces.
2. Pin LIKE. A fourth DuckDB-generated matrix set (EXPECTED_LIKE), which records six wrong WHERE answers. Three corpus cases over new data/like_patterns.csv pin it end to end (DIFFER, DIFFER, GAP).
3. Fix the exponential LIKE matcher. The value evaluator's matcher was right on all ten patterns but recursive: one 80-character value against '%a%a%a%a%a%a%a%b' ran for over a minute (SELECT has had this all along). Replaced with the iterative wildcard algorithm, O(text × pattern), before WHERE switched to it. Adds 28 unit cases plus a 2,000-character case with a 1 s budget; the CLI case now takes 57 ms.
4. LIKE over a number: D2. DuckDB rejects it, so there's no reference answer. Decision: match the number's displayed text, as the value evaluator already did (5.0 displays as 5). Recorded in SQL_PARITY.md as a deliberate difference from DuckDB. Pinned as a fifth matrix set whose expected answers are this decision; WHERE's old "always FALSE" is recorded against it.
5. Delegate LIKE. All of LIKE (NULL handling, the text of a number, case mode, matching) is now one function, sql_like, called by the value evaluator and by WHERE's one remaining legacy arm (a method call on the left). Exactly the ten pinned WHERE answers changed, and parity moved by exactly the three P55 cases: 176 → 179 AGREE / 209.
6. Remove the plumbing. Deletes EvaluationContext, WHERE's 'ctx lifetime and context constructors, with_config (no callers) and the always-zero cache stats. 147 lines out, parity byte-identical.
7. Docs: R13 entry, P55 and D2 entries, headline counts.

Verification

• Parity: before and after reports compared at each commit. Nothing moved except the pinned cases; contract holds at 209.
• Evaluator matrix green, all six sets.
• ./run_all_tests.sh green: 830 lib + 501 integration Rust tests, 572 Python tests, all FORMAL examples. The one SMOKE failure (expander_rewriters query 7) also fails on main.
• Over trades_100000.csv, WHERE IS NULL and both LIKE shapes stay at CSV load time without the regex cache.

Next: slice 5, method calls

WHERE's own Contains, StartsWith and EndsWith always ignore case and answer FALSE for NULL, and it has its own left-side Length, IndexOf and Trim*. These need pinning against the registry functions the value evaluator maps them to, then moving over.

🤖 Generated with Claude Code